AI-based detection of thrips and whitefly species on sticky traps in greenhouses and vertical farms

In 2021, the Flemish ornamental horticulture sector boasted a production value exceeding 500 million euros, focused on top quality products predominantly aimed at export markets. As growers aim to increase sustainability and reduce resource dependency in terms of energy, land, water and nutrients, the switch to more efficient and sustainable systems such as ornamental vertical farming (VF) will be needed. Unfortunately, the presence of pest insects and the transmission of insect-borne diseases can potentially ruin complete cultivations, especially inside VF systems where both plants and pests can flourish due to the optimal indoor climate. As the risk of direct crop damage and/or disease transmission strongly depends on the abundance and pest insect species, early detection, identification and fast/local treatment are key. In greenhouses nowadays, the pest scouting process is generally performed by spreading sticky traps across the compartments, followed by frequent human inspection. Although this method is relatively cheap and effective, it will not be feasible in large greenhouse facilities or multilayer VF systems due to the high need for human interaction. Therefore, the objective of this study is to develop a pest detection and identification system at species level using sticky trap images and deep learning (DL) for both greenhouse and VF settings.

As use cases, we selected two species of thrips: Frankliniella occidentalis and Echinothrips americanus, along with two species of whiteflies: Bemisia tabaci and Trialeurodes vaporariorum. In order to easily obtain the labelled pest images on sticky traps, each pest species was reared inside insect-proof cages in greenhouses and growing chambers of ILVO1 and PCS2. Subsequently, various types of yellow sticky traps (YSTs) were positioned inside the insect cages for several days and afterwards photographed inside a dedicated photography chamber using a high resolution DSLR camera. After image correction, all insect images were subsequently semi-automatically labelled using a dedicated Python script and manually verified. Currently, various object detection and classification models are being trained and tested using the labelled images in order to evaluate the potential of DL-based species-level detection of thrips and whitefly on YSTs for both greenhouse and VF settings.

Future research will encompass a) expanding the system to incorporate additional pest species and types of sticky traps, b) testing the models across different light conditions and c) determining the minimal required image resolution for detection of pests at the species level.

Keywords: integrated pest management (IPM), deep learning (DL), automatic pest detection, automatic insect monitoring, YSTs

Acknowledgement: This research is part of the VLAIO-LA project ‘Ornamental cultivation moves to a higher level’ (2022-2026) and is funded by the Flemish government and EU Blue Deal.